A PROBABILITY MODEL FOR GOLF

PUTTING

AJ Clair, Tommy Durand, & Jeremy Polster

Golf Background Golf is hard! Some view putting as the most difficult

part of golf A study examining professional golfers

showed that they were successful on less than 60% of their five-foot putts.

Putting Data Berry, D. Statistics: A Bayesian Perspective. Belmont, CA. 1995: Duxbury Press Distance (ft) Tried Success % Made 2 1443 1346 93.2800 3 694 577 83.1400 4 455 337 74.0700 5 353 208 58.9200 6 272 149 54.7800 7 256 136 53.1200 8 240 111 46.2500 9 217 69 31.8000 10 200 67 33.5000 11 237 75 31.6500 12 202 52 25.7400 13 192 46 23.9600 14 174 54 31.0300 15 167 28 16.7700 16 201 27 13.4300 17 195 31 15.9000 18 191 33 17.2800 19 147 20 13.6100 20 152 24 15.7900

Looking at an Appropriate Putting Model

We can view this as a trigonometric model When Ө is less than the “threshold angle” the ball

will go in the hole. Threshold angle = arcsin[(R-r)/x]

Can the Model Be Used to Estimate Probability of Successful Putts?

We can imagine a normal distribution to represent the random variable, Ө

Probability Model Using this normal distribution, the

probability of a successful putt is:

Φ represents the standard normal distribution function and σ is the standard deviation

The hole diameter, 2R = 4.25 inches and the ball diameter, 2r = 1.68 inches

Using several mathematical properties including the small angle approximation for the arcsin, we can simplify the equation to:

Using this approximation we calculate the probabilities of a putt at all the given distances. We do this using the estimated value for σ = 0.026 (1.5 degrees)

Distance (ft) Success Probability (Estimated)

6 0.5476767 0.4694378 0.4107579 0.365117

10 0.32860611 0.29873212 0.27383813 0.25277414 0.23471815 0.2190716 0.20537917 0.19329718 0.18255919 0.1729520 0.164303

Problems with Linear and Quadratic Models

Linear2 to 3 ft change is not the same as 19 to 20

ft changeProbability model will show probabilities not

bound between 0 and 1 Quadratic

Might have larger R-sq yet an increasing probability at higher distance is not likely

Also, can show probabilities not bound between 0 and 1

Alternative Probability Model One common estimator for probability

models is a d-probit estimator This is a logistic function which is bound

between 0 and 1

Our Experiment Using the actual probabilities for putting

given by Berry, we simulated 100 putts for each distance using binomial random data

We then ran a d-probit regression on the data

Our Results The coefficient for distance in our d-

probit regression was -.026567 This is interpreted as a one foot

increase in distance results in a 2.66% decrease in the probability of success for a putt

The 95% confidence interval for the coefficient was (-.031888, -.021246)

Discussion Note both linear regressions for the

actual probabilities and the equation estimated probabilities have slopes included in the confidence interval for our experiment’s coefficient

These slopes are interpreted the same way as the coefficient

Problems With Model The model only accounts for the angle

at which the ball is hit Does not account for putts that are too

short and balls which partly cover the hole and go in

Also it only accounts for distance not terrain, playing conditions, or golfer ability

Conclusion We can conclude that for our data range,

our linear probability models are quite accurate at predicting the change in probability from the change in distance.

Also, the equation for estimating the probabilities appears to be a good estimate for the actual probabilities

However with a larger range of data the linear probability model may not be suitable

References Berry, D. Statistics: A Bayesian

Perspective. Belmont, CA. 1995: Duxbury Press

http://godplaysdice.blogspot.com/2008/06/probability-of-making-putt.html

Gelman, A., Nolan D. “A Probability Model for Golf Putting.” Teaching Statistics Vol. 24 No. 3 pg. 93-95